Book

Every practitioner who has maintained a machine learning pipeline has a story like this: a routine library upgrade, no errors, no warnings — and yet the model's output quietly changed. This is the story of default-argument breaking changes (DABCs): silent library updates that alter a function's behavior without ever raising an exception. This book introduces a differential-testing methodology that measures, for the first time, how severely these silent changes actually affect real output — not just whether client code happens to call the changed function. Applying it to a catalog of 88 documented breaking changes across scikit-learn and pandas, the author verified 13 directly, and used the results to train an explainable classifier that predicts severity for the rest with 97% cross-validated accuracy. Along the way, the book uncovers a precisely quantified limitation of differential testing itself: 85% of an eight-year breaking-change catalog can no longer be dynamically verified, due to the disappearance of compatible Python interpreters — a finding with implications far beyond this one study. Written for software-evolution researchers and working ML practitioners alike, this book pairs rigorous methodology with a practical, deployable tool: SilentDrift, which lets any developer check their exposure before the next dependency upgrade.